Anthropic Mythos 5.2 ExploitBench - Plain Mean Flags: leaderboard

Anthropic's final Claude Mythos 5.2 evaluation on 41 V8 vulnerabilities, with 16 mechanically verified exploitation capability flags, five trials per vulnerability and a 300-turn budget. The plain arm makes uninterrupted attempts in the authors' static uniform harness, with cyber safeguards disabled and sandbox escape classifiers enabled. This metric averages flags per trial over all environments on a 0–16 scale. Reasoning effort is unreported; no plain-arm comparator values are published. AutoNudge is a separate four-model evaluation.

Metric: Mean capability flags per trial (plain arm; 0–16). Source: assets.claude.com. 1 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=anthropic-mythos-5-2-exploitbench-plain-mean-flags · How It Works · Data refreshed daily, snapshot 2026-11-12.